1. Field of the Invention
The present invention generally relates to the manufacturing of metal-oxide-semiconductor field effect transistors (MOSFETs), and more particularly to MOSFETs manufactured for reproducibility of threshold voltages among otherwise identical transistors.
2. Prior Art
Random variation in threshold voltage (σVT) of metal-oxide semiconductor (MOS) field effect transistors (MOSFETs) with high-K (high dielectric constant) metal gate stack is caused by some dominant factors: (i) random dopant fluctuations (RDF) in the well and in the pocket implant regions underneath the gate, which, among other things, cause variations in depletion layer thickness; (ii) line edge roughness (LER) which causes random variation in the length of the gate electrode resulting from random variations in profile of the etched gate; and, (iii) metal gate granularity (MGG) which causes random variations in the local work function due to the grain structure of the gate material. There is a fourth source of variation, the random variations in the effective channel length, referred to as random extension fluctuations (RXF), arising from statistical variations in the position of the junction that separates the channel from either the source or the drain extensions. However, as MOSFETs become smaller, the effects of RDF, LER, and RXF increase and become major factors in determining σVT. The first effect, RDF, has recently gained intense attention. The randomness in the position of the drain extension RXF has two principal sources: a) variations in the final position of implanted ions due to scattering; and, b) variations in the activation and positions of the source/drain extension ions as influenced by the activation and subsequent heat treatments.
It is well-known in the art that as MOSFETs move to finer and finer dimensions, variability of the threshold voltage σVT seriously undermines the reproducibility of threshold voltages among otherwise identical transistors. This effect is inevitable, and it is particularly severe in its impact on complimentary MOS (CMOS) static random access memories (SRAM), which use millions of near-minimum sized transistors. Development of ultra-thin silicon on insulator (SOI) structures, e.g., fully depleted SOI (FDSOI), and of three-dimensional transistors (FinFET and Tri-Gate), are largely motivated by a need to reduce the threshold spreads σVT caused by RDF. This trend moves away from the more traditional bulk MOS manufacturing, adversely impacting costs and availability. A cross-section 400 of a standard bulk MOSFET, formed in a gate-last process, is shown in FIG. 4. On bulk 410 of one conductivity type, drain and source areas 420 of an opposite conductivity type are formed. A SiO2 isolation layer 430 is formed over the entire MOSFET transistor, with openings for connections 470 to respective drain and source terminals. In the gate-last process of the MOSFET, the gate is formed by having spacers 440 formed above the SiO2 layer. The SiO2 is removed and typically replaced by a high-K dielectric stack 450 on top of which a metal gate 460 is formed. In certain embodiments the spacers are formed by repeating the deposition-and-etch process.
One method of addressing the RDF problem was described by Asenov et al. in the paper “Suppression of Random Dopant-Induced Threshold Voltage Fluctuations in Sub-0.1-μm MOSFETs with Epitaxial and δ-Doped Channels,” IEEE Transactions on Electron Devices, Vol. 46, No. 8, August 1999, Pages 1718-1724. This approach is consistent with bulk transistors, and does not have the same cost penalties associated with FinFETs and FDSOI. This scheme has three key components: a) placing a thin, approximately 10 nanometer (nm), minimally doped epitaxial layer immediately beneath the gate oxide; b) placing a thinner layer with a very high concentration of either acceptors for an NMOS device or donors for a PMOS device at the boundary of the thin epitaxial layer that is remote from the gate dielectric interface; and, c) incorporating a moderately heavily doped well layer beneath the un-doped epitaxial layer and the highly doped, thinner layer. Similar structures have been demonstrated by Fujita et al. as described in their paper “Advanced Channel Engineering Achieving Aggressive Reduction of VT Variation for Ultra-Low-Power Applications”, Electron Devices Meeting (IEDM), 2011 IEEE International, pp. 32.3.1-32.3.4, 5-7 Dec. 2011. A cross-section 500 of such an epitaxial transistor is shown in FIG. 5. The epitaxial layer 510 is deposited on the whole silicon wafer after the channel doping implantation through a sacrificial gate oxide and before the formation of the sacrificial polysilicon gate and the source drain regions 525. In some instances, for example, Hokazono, A., et al., in “25-nm Gate Length nMOSFET With Steep Channel Profiles Utilizing Carbon-Doped Silicon Layers (A P-Type Dopant Confinement Layer)”, Electron Devices, IEEE Transactions on Electron Devices, vol. 58, no. 5, pp. 1302-1310, May 2011, carbon is introduced before the epitaxy to retard the diffusion of doping into the low-doped epitaxial region during the source/drain implantation activation. However, this is an exceedingly difficult task and experimentation has shown that the profiles degrade due to subsequent processing steps, resulting in doping penetration into the low-doped epitaxial layer.
Roy, et al. in “Random Dopant Fluctuation Resistant ‘Bulk’ MOSFETs with Epitaxial Delta Doped Channels,” Ultimate Integration in Silicon (ULIS) Conference, Glasgow, Scotland, 2007, further amplified these concepts when addressing a thin epitaxial layer, plus 6-layer in the context of a known 45 nm CMOS technology, having a gate length of 35 nm. In each case the critical δ-layer doping was simulated as if it had a depth distribution of a Dirac delta function, i.e., the doping ions are distributed in a differentially thin layer. In all cases, substantial reductions in the statistical variations in threshold voltage σVT due to RDF were realized. The delta doping layer, in these simulations, is mostly depleted, and that layer provides the tool for adjusting the nominal value of the threshold voltage. However, high temperature processing would substantially alter the doping profiles, compromising the expected benefits. This problem persists even though it is well known that certain complementary dopants, like carbon, can retard motion during thermal treatments, at least in n-channel devices.
Recent changes to gate architectures from poly-silicon over silicon dioxide or oxynitride, to metal over high-K gate dielectric stack have changed the processing sequence. (A high-K or high dielectric constant as used herein and in the claims to follow means a dielectric constant that is higher than the dielectric constant of silicon dioxide (K=3.9); an effective dielectric constant K exceeding 6 would be a preferred high-K dielectric.) This is particularly true for the “gate last” process flow. In this process, while there is seemingly a complete transistor, with oxide and/or nitride sidewall spacers on each side of a poly gate, the gate is actually a sacrificial structure. That gate structure and the underlying protective oxide are etched away, exposing the silicon surface. Then a sequence of steps is employed: a) deposition of a high-K gate dielectric, typically by atomic layer deposition; b) deposition of a metal gate having a controlled work function to set the threshold voltage; and, c) deposition of a robust gate material, typically doped amorphous silicon. In many cases the sequence above is augmented by chemical-mechanical polishing steps to assure the localization of the various layers. However, this structure does not overcome the deficiencies resulting from RDF, LER or RXF. The morphology of the metal in the gate last process does reduce the variations identified as MGG, compared to a gate first process.
It would therefore be advantageous to find a solution compatible with bulk MOS manufacturing that would overcome deficiencies resulting from RDF, RXF and LER while generally maintaining the cost advantages and relative simplicity of standard bulk MOS manufacturing. It would be further advantageous if reductions of RDF, RXF, and LER are achieved simultaneously.